Instruction following issues
Hello,
When it comes to writing stories, the 70B version is amazing, but this is even better. I am running the Q6_K quant on a 2XA40 in runpod.
However, when I give it long instructions, it eventually starts to go off on its own and forget the later parts of my instructions. For example, here is my prompt.
"Write a story about a dramatic air battle between the U.S. and Russian air forces, showcasing the perspectives of both sides. Use the following details for the disposition of forces and the sequence of events:
American Forces:
• Four aircraft: Eagle1, Eagle2, Eagle3, Eagle4
• Eagle4 is the leader
Russian Forces:
• Four aircraft: Tiger1, Tiger2, Tiger3, Tiger4
• Tiger4 is the leader
Battle Outline: you should follow this outline exactly:
Note: The battle is long and brutal, with destroyed aircraft meaning the pilot is killed.
- Eagle3 attacks Tiger1. Tiger1 is destroyed.
- Eagle4 attacks Tiger2. Tiger2 is destroyed.
- Tiger4 attacks Eagle1. Eagle1 is destroyed.
- Tiger4 attacks Eagle2. Eagle2 is destroyed.
- Eagle3 attacks both Tiger3 and Tiger4, while Tiger3 attacks Eagle3. Eagle3 and Tiger3 are destroyed, and Tiger4 is heavily damaged.
- Eagle4 and Tiger4 attack each other. Eagle4 is destroyed, and Tiger4 is heavily damaged."
Now, every time I feed this prompt, the model executes outlines 1-4 perfectly, but falls flat on 5 and 6. Particularly on 6, sometimes both Tiger4 and Eagle4 are destroyed, other times Tiger4 is the one destroyed.
Perhaps the model is not that good on instruction following, but I believe what's more likely is I am prompting it incorrectly.
If it is alright with you, can you help me out on this? I usually do prompts like these, I ask the model to create the story, and then provide the outline of the plot. For small outlines, the model works fine, but for longer instructions like the above, it starts to not follow the outline near the end.
Thank you!
Every model has its weaknesses. That prompt is just detailed enough across enough discrete steps that I wouldn't expect stellar results from a LLM from early 2024. I don't think you're doing anything wrong with your prompting. It's more like you've found the limit of what the model is capable of doing.
Ah, perhaps your right. I tried my prompt on both Command R+ and Mistral-Large-Instruct-2407 a few hours ago. Both followed my outline without mistakes, but the writing is not good at all. This made me curious. Is there a tradeoff between smarts vs creativeness?
American Forces:
• Four aircraft: Eagle1, Eagle2, Eagle3, Eagle4
• Eagle4 is the leader
Russian Forces:
• Four aircraft: Tiger1, Tiger2, Tiger3, Tiger4
• Tiger4 is the leader
It might work better if you give different names instead of adding a number to the end.
I've found that merged models often struggle more with multi-character POVs in general though sadly.